Revert "Merge pull request #196 from edx/platform/upgrade-to-ruby23"

This reverts commit 22178086, reversing
changes made to 14ed8a85.

Upgrade to Ruby 2.3.0

Use proper relations instead of forcefully loading relations by hand.

Some of the ways comments are accessed in the User model (when a
complete user object is requested) lead to an N+1 situation because
while we preload the related documents (preloading the parent thread of
  a given comment, in this case), further accesses to those comments
  actually never take advantage of the preloaded relations.

We should be using the relation directly, which will still act the same
as before but allow us to actually benefit from preloading.

Wait 10s after starting ElasticSearch.

Now that MongoDB 3.0 seems to download/untar faster than 2.6, we seem to
be right over the line of how quickly ES can start up and be listening
for connections, which leads the tests to try and connect too quickly,
which fails.

By sleeping for 10 seconds after starting ElasticSearch, it's almost a
certainty it will be up and running by the time we try to first connect
to it.

Switch to using MongoDB 3.0 in tests.

Now that we're on MongoDB 3.0 in production, we should also be testing
with MongoDB 3.0.

Add message for rack config load.

Fix new relic traces.

Revert "Replaced Tire with elasticsearch-model"

This reverts commit 56b61a13.

Replaced Tire with elasticsearch-model

Updated the Mongoid configuration options

Updated the Mongoid configuration so that the read and write modes can
be specified via the `forum_env`.

Reduced merges, do them in place, and reduce allocations.

update newrelic

Primarily, we wanted to reduce the one-thing-per-merge calls, but also,
we want to reduce allocations overall, both from merge calls which will
duplicate the source hash and from string constants sprinkled
throughout.

Seemingly, defining a constant-style word list is not the ticket.  For
whatever reason, Ruby can't seem to figure out the list doesn't change,
and presumably is splating out all of the words and allocating them as
if they were declared right there.  Wahhhh.

Now we're using a straight up one-for-one constant definition of each
string, in a big constants file, where we freeze the strings on the
spot.  We then reference each constant directly at the callsite, which
was the original style and, ultimately, is the best for readability.

Handle comment context in cases of missing parent thread.

TNL-4944

Upgrade nokogiri to address security issues in libxml2

Upgrade nokogiri to address security issues in libxml2 https://github.com/sparklemotion/nokogiri/issues/1473